LoopVQ crushes KV cache to 0.25 bits

🔥 What happened Researchers from Seoul National University dropped LoopVQ, a KV-cache quantization method for looped transformers that hits 0.5 bits per element — and 0.25 bits when combined with TaSQ. Implemented in SGLang, it nearly halves decode latency at 32K context and boosts peak throughput up to 6.1x on Ouro models. 💡 Why it matters Looped models like Ouro and Huginn save parameters by reusing layers, but their KV cache grows linearly with each loop step — the real memory bottleneck. LoopVQ exploits redundancy along the loop axis and quantizes multiple steps jointly. At 0.25 bits, it retains much of BF16 accuracy even on long chain-of-thought reasoning. ⚡ Our take Finally, someone treats the loop cache as the memory leak it is. If you're deploying looped transformers today, LoopVQ is not optional — 6x throughput isn't a benchmark trick, it's real money.